How Google Builds an AI Overview, From Your Question to the Answer on Screen

When you type a question into Google and a paragraph of AI-written text appears above the blue links, several things have already happened in under two seconds. Google decides the query deserves a generated answer, pulls from pages it has already indexed, drafts a response with a language model, and attaches source links. You are reading a summary, not a search result.

That distinction matters more than most people realize, and it changes how a business gets found. Below is the actual sequence, step by step, with the parts that surprise people flagged as they come up.

Step one: Google decides whether your question earns an answer

Not every search triggers an AI Overview. Google runs the query through a classifier that predicts whether a synthesized answer will help more than a plain list of links. Informational and how-to questions often qualify. Navigational searches, where you already know the site you want, usually do not. The decision happens before any answer is written.

You can watch this yourself. Search “what does a cardiologist treat” and an Overview is likely. Search the name of a specific clinic and you will probably just get links, because Google reads that as intent to visit one known destination. The system is guessing at what you meant, and it guesses conservatively on anything where a wrong summary could mislead. Health queries get more of that caution than most categories, which is Google’s own stated position under its “Your Money or Your Life” framework for topics that affect safety and wellbeing.

Step two: it retrieves passages, not whole pages

Once a query qualifies, Google gathers candidate material from its existing index. It does not read the open web live for each search. It pulls passages it has already crawled and stored, then ranks them for relevance to your specific wording. This is why a page that was never indexed cannot appear in an Overview, and why fresh content sometimes lags.

The unit here is the passage, not the page. Google frequently lifts a self-contained two or three sentence answer from deep inside an article rather than the whole thing. A page that answers a clear question directly, in a block that stands on its own, is easier to quote than one where the answer is scattered across five paragraphs. That single fact reshapes how pages worth citing get written.

Step three: a language model drafts the summary

Google’s model then composes the visible answer from the retrieved passages. It is not writing from memory or inventing prose. It is stitching and paraphrasing sourced material, a method Google describes as grounding, meant to keep the output tied to real pages rather than to whatever the model might otherwise generate. Grounding reduces fabrication. It does not eliminate it.

This is the step people trust too much. The draft can still blend two sources into a claim neither one actually made, or carry forward an error that was already sitting in a ranked page. For medical questions that risk is not academic, and it is exactly where practices should care about which of their pages Google is drawing from. If you run a clinic and want to understand how your own content shapes what Google says about your specialty, PracticeRank has more on google ai overviews and how they pull from medical sites. Treat the generated paragraph as a starting point, not a citation, and verify anything clinical against a named authority.

Step four: it attaches sources and renders the card

Before the answer reaches your screen, Google links the passages it used. Those source cards on the right or below the text are not decoration. They are the pages that fed the summary, and clicking through is how you check whether the Overview actually reflects them. When the links and the text disagree, the links are the record.

The rendering is fast because most of the heavy work is precomputed. Retrieval runs against an index that already exists. The model draft for common queries can be cached and reused, so two people asking the same thing minutes apart may see identical wording. What feels like a live answer is often a stored one, refreshed on Google’s schedule rather than yours.

What this changes for anyone who wants to be cited

Being quoted in an Overview is not the same as ranking first, and chasing the old goal can cost you the new one. Google is selecting passages that answer a question cleanly and come from a page it already trusts on that topic. Length and keyword repetition do not help. A direct, sourced answer near the top of a relevant page does.

Here is the position worth taking: AI Overviews reward pages that behave like a good reference, not pages built to game a ranking. That means writing the answer to a real question in the first few lines, backing specifics with a source a reader can open, and covering a subject you have genuine standing to discuss. A thin page stuffed with the right words used to work. It does not get lifted into an Overview, because the model has nothing quotable to lift.

The practical takeaway is smaller than the anxiety around it. Learn which of your questions trigger an Overview, read the sources Google chose, and notice what those pages did that yours did not. The machine is showing its work in those source cards. Most people never look.